/* * Copyright 2012 Advanced Micro Devices, Inc. * * SPDX-License-Identifier: MIT */ #include "ac_shader_util.h" #include "ac_gpu_info.h" #include "amdgfxregs.h" #include "util/u_math.h" #include unsigned ac_get_spi_shader_z_format(bool writes_z, bool writes_stencil, bool writes_samplemask, bool writes_mrt0_alpha) { /* RGBA = (Z, stencil, samplemask, mrt0_alpha). * Both stencil and sample mask need only 16 bits. */ if (writes_mrt0_alpha) { if (writes_stencil || writes_samplemask) return V_028710_SPI_SHADER_32_ABGR; else return V_028710_SPI_SHADER_32_AR; } if (writes_samplemask) { if (writes_z) return V_028710_SPI_SHADER_32_ABGR; else return V_028710_SPI_SHADER_UINT16_ABGR; } if (writes_stencil) return V_028710_SPI_SHADER_32_GR; else if (writes_z) return V_028710_SPI_SHADER_32_R; else return V_028710_SPI_SHADER_ZERO; } unsigned ac_get_cb_shader_mask(unsigned spi_shader_col_format) { unsigned i, cb_shader_mask = 0; /* If the format is ~0, it means we want a full mask. */ if (spi_shader_col_format == ~0) return ~0; for (i = 0; i < 8; i++) { switch ((spi_shader_col_format >> (i * 4)) & 0xf) { case V_028714_SPI_SHADER_ZERO: break; case V_028714_SPI_SHADER_32_R: cb_shader_mask |= 0x1 << (i * 4); break; case V_028714_SPI_SHADER_32_GR: cb_shader_mask |= 0x3 << (i * 4); break; case V_028714_SPI_SHADER_32_AR: cb_shader_mask |= 0x9u << (i * 4); break; case V_028714_SPI_SHADER_FP16_ABGR: case V_028714_SPI_SHADER_UNORM16_ABGR: case V_028714_SPI_SHADER_SNORM16_ABGR: case V_028714_SPI_SHADER_UINT16_ABGR: case V_028714_SPI_SHADER_SINT16_ABGR: case V_028714_SPI_SHADER_32_ABGR: cb_shader_mask |= 0xfu << (i * 4); break; default: assert(0); } } return cb_shader_mask; } /** * Calculate the appropriate setting of VGT_GS_MODE when \p shader is a * geometry shader. */ uint32_t ac_vgt_gs_mode(unsigned gs_max_vert_out, enum amd_gfx_level gfx_level) { unsigned cut_mode; assert (gfx_level < GFX11); if (gs_max_vert_out <= 128) { cut_mode = V_028A40_GS_CUT_128; } else if (gs_max_vert_out <= 256) { cut_mode = V_028A40_GS_CUT_256; } else if (gs_max_vert_out <= 512) { cut_mode = V_028A40_GS_CUT_512; } else { assert(gs_max_vert_out <= 1024); cut_mode = V_028A40_GS_CUT_1024; } return S_028A40_MODE(V_028A40_GS_SCENARIO_G) | S_028A40_CUT_MODE(cut_mode) | S_028A40_ES_WRITE_OPTIMIZE(gfx_level <= GFX8) | S_028A40_GS_WRITE_OPTIMIZE(1) | S_028A40_ONCHIP(gfx_level >= GFX9 ? 1 : 0); } /// Translate a (dfmt, nfmt) pair into a chip-appropriate combined format /// value for LLVM8+ tbuffer intrinsics. unsigned ac_get_tbuffer_format(enum amd_gfx_level gfx_level, unsigned dfmt, unsigned nfmt) { // Some games try to access vertex buffers without a valid format. // This is a game bug, but we should still handle it gracefully. if (dfmt == V_008F0C_GFX10_FORMAT_INVALID) return V_008F0C_GFX10_FORMAT_INVALID; if (gfx_level >= GFX11) { switch (dfmt) { default: UNREACHABLE("bad dfmt"); case V_008F0C_BUF_DATA_FORMAT_INVALID: return V_008F0C_GFX11_FORMAT_INVALID; case V_008F0C_BUF_DATA_FORMAT_8: switch (nfmt) { case V_008F0C_BUF_NUM_FORMAT_UNORM: return V_008F0C_GFX11_FORMAT_8_UNORM; case V_008F0C_BUF_NUM_FORMAT_SNORM: return V_008F0C_GFX11_FORMAT_8_SNORM; case V_008F0C_BUF_NUM_FORMAT_USCALED: return V_008F0C_GFX11_FORMAT_8_USCALED; case V_008F0C_BUF_NUM_FORMAT_SSCALED: return V_008F0C_GFX11_FORMAT_8_SSCALED; default: UNREACHABLE("bad nfmt"); case V_008F0C_BUF_NUM_FORMAT_UINT: return V_008F0C_GFX11_FORMAT_8_UINT; case V_008F0C_BUF_NUM_FORMAT_SINT: return V_008F0C_GFX11_FORMAT_8_SINT; } case V_008F0C_BUF_DATA_FORMAT_8_8: switch (nfmt) { case V_008F0C_BUF_NUM_FORMAT_UNORM: return V_008F0C_GFX11_FORMAT_8_8_UNORM; case V_008F0C_BUF_NUM_FORMAT_SNORM: return V_008F0C_GFX11_FORMAT_8_8_SNORM; case V_008F0C_BUF_NUM_FORMAT_USCALED: return V_008F0C_GFX11_FORMAT_8_8_USCALED; case V_008F0C_BUF_NUM_FORMAT_SSCALED: return V_008F0C_GFX11_FORMAT_8_8_SSCALED; default: UNREACHABLE("bad nfmt"); case V_008F0C_BUF_NUM_FORMAT_UINT: return V_008F0C_GFX11_FORMAT_8_8_UINT; case V_008F0C_BUF_NUM_FORMAT_SINT: return V_008F0C_GFX11_FORMAT_8_8_SINT; } case V_008F0C_BUF_DATA_FORMAT_8_8_8_8: switch (nfmt) { case V_008F0C_BUF_NUM_FORMAT_UNORM: return V_008F0C_GFX11_FORMAT_8_8_8_8_UNORM; case V_008F0C_BUF_NUM_FORMAT_SNORM: return V_008F0C_GFX11_FORMAT_8_8_8_8_SNORM; case V_008F0C_BUF_NUM_FORMAT_USCALED: return V_008F0C_GFX11_FORMAT_8_8_8_8_USCALED; case V_008F0C_BUF_NUM_FORMAT_SSCALED: return V_008F0C_GFX11_FORMAT_8_8_8_8_SSCALED; default: UNREACHABLE("bad nfmt"); case V_008F0C_BUF_NUM_FORMAT_UINT: return V_008F0C_GFX11_FORMAT_8_8_8_8_UINT; case V_008F0C_BUF_NUM_FORMAT_SINT: return V_008F0C_GFX11_FORMAT_8_8_8_8_SINT; } case V_008F0C_BUF_DATA_FORMAT_16: switch (nfmt) { case V_008F0C_BUF_NUM_FORMAT_UNORM: return V_008F0C_GFX11_FORMAT_16_UNORM; case V_008F0C_BUF_NUM_FORMAT_SNORM: return V_008F0C_GFX11_FORMAT_16_SNORM; case V_008F0C_BUF_NUM_FORMAT_USCALED: return V_008F0C_GFX11_FORMAT_16_USCALED; case V_008F0C_BUF_NUM_FORMAT_SSCALED: return V_008F0C_GFX11_FORMAT_16_SSCALED; default: UNREACHABLE("bad nfmt"); case V_008F0C_BUF_NUM_FORMAT_UINT: return V_008F0C_GFX11_FORMAT_16_UINT; case V_008F0C_BUF_NUM_FORMAT_SINT: return V_008F0C_GFX11_FORMAT_16_SINT; case V_008F0C_BUF_NUM_FORMAT_FLOAT: return V_008F0C_GFX11_FORMAT_16_FLOAT; } case V_008F0C_BUF_DATA_FORMAT_16_16: switch (nfmt) { case V_008F0C_BUF_NUM_FORMAT_UNORM: return V_008F0C_GFX11_FORMAT_16_16_UNORM; case V_008F0C_BUF_NUM_FORMAT_SNORM: return V_008F0C_GFX11_FORMAT_16_16_SNORM; case V_008F0C_BUF_NUM_FORMAT_USCALED: return V_008F0C_GFX11_FORMAT_16_16_USCALED; case V_008F0C_BUF_NUM_FORMAT_SSCALED: return V_008F0C_GFX11_FORMAT_16_16_SSCALED; default: UNREACHABLE("bad nfmt"); case V_008F0C_BUF_NUM_FORMAT_UINT: return V_008F0C_GFX11_FORMAT_16_16_UINT; case V_008F0C_BUF_NUM_FORMAT_SINT: return V_008F0C_GFX11_FORMAT_16_16_SINT; case V_008F0C_BUF_NUM_FORMAT_FLOAT: return V_008F0C_GFX11_FORMAT_16_16_FLOAT; } case V_008F0C_BUF_DATA_FORMAT_16_16_16_16: switch (nfmt) { case V_008F0C_BUF_NUM_FORMAT_UNORM: return V_008F0C_GFX11_FORMAT_16_16_16_16_UNORM; case V_008F0C_BUF_NUM_FORMAT_SNORM: return V_008F0C_GFX11_FORMAT_16_16_16_16_SNORM; case V_008F0C_BUF_NUM_FORMAT_USCALED: return V_008F0C_GFX11_FORMAT_16_16_16_16_USCALED; case V_008F0C_BUF_NUM_FORMAT_SSCALED: return V_008F0C_GFX11_FORMAT_16_16_16_16_SSCALED; default: UNREACHABLE("bad nfmt"); case V_008F0C_BUF_NUM_FORMAT_UINT: return V_008F0C_GFX11_FORMAT_16_16_16_16_UINT; case V_008F0C_BUF_NUM_FORMAT_SINT: return V_008F0C_GFX11_FORMAT_16_16_16_16_SINT; case V_008F0C_BUF_NUM_FORMAT_FLOAT: return V_008F0C_GFX11_FORMAT_16_16_16_16_FLOAT; } case V_008F0C_BUF_DATA_FORMAT_32: switch (nfmt) { default: UNREACHABLE("bad nfmt"); case V_008F0C_BUF_NUM_FORMAT_UINT: return V_008F0C_GFX11_FORMAT_32_UINT; case V_008F0C_BUF_NUM_FORMAT_SINT: return V_008F0C_GFX11_FORMAT_32_SINT; case V_008F0C_BUF_NUM_FORMAT_FLOAT: return V_008F0C_GFX11_FORMAT_32_FLOAT; } case V_008F0C_BUF_DATA_FORMAT_32_32: switch (nfmt) { default: UNREACHABLE("bad nfmt"); case V_008F0C_BUF_NUM_FORMAT_UINT: return V_008F0C_GFX11_FORMAT_32_32_UINT; case V_008F0C_BUF_NUM_FORMAT_SINT: return V_008F0C_GFX11_FORMAT_32_32_SINT; case V_008F0C_BUF_NUM_FORMAT_FLOAT: return V_008F0C_GFX11_FORMAT_32_32_FLOAT; } case V_008F0C_BUF_DATA_FORMAT_32_32_32: switch (nfmt) { default: UNREACHABLE("bad nfmt"); case V_008F0C_BUF_NUM_FORMAT_UINT: return V_008F0C_GFX11_FORMAT_32_32_32_UINT; case V_008F0C_BUF_NUM_FORMAT_SINT: return V_008F0C_GFX11_FORMAT_32_32_32_SINT; case V_008F0C_BUF_NUM_FORMAT_FLOAT: return V_008F0C_GFX11_FORMAT_32_32_32_FLOAT; } case V_008F0C_BUF_DATA_FORMAT_32_32_32_32: switch (nfmt) { default: UNREACHABLE("bad nfmt"); case V_008F0C_BUF_NUM_FORMAT_UINT: return V_008F0C_GFX11_FORMAT_32_32_32_32_UINT; case V_008F0C_BUF_NUM_FORMAT_SINT: return V_008F0C_GFX11_FORMAT_32_32_32_32_SINT; case V_008F0C_BUF_NUM_FORMAT_FLOAT: return V_008F0C_GFX11_FORMAT_32_32_32_32_FLOAT; } case V_008F0C_BUF_DATA_FORMAT_2_10_10_10: switch (nfmt) { case V_008F0C_BUF_NUM_FORMAT_UNORM: return V_008F0C_GFX11_FORMAT_2_10_10_10_UNORM; case V_008F0C_BUF_NUM_FORMAT_SNORM: return V_008F0C_GFX11_FORMAT_2_10_10_10_SNORM; case V_008F0C_BUF_NUM_FORMAT_USCALED: return V_008F0C_GFX11_FORMAT_2_10_10_10_USCALED; case V_008F0C_BUF_NUM_FORMAT_SSCALED: return V_008F0C_GFX11_FORMAT_2_10_10_10_SSCALED; default: UNREACHABLE("bad nfmt"); case V_008F0C_BUF_NUM_FORMAT_UINT: return V_008F0C_GFX11_FORMAT_2_10_10_10_UINT; case V_008F0C_BUF_NUM_FORMAT_SINT: return V_008F0C_GFX11_FORMAT_2_10_10_10_SINT; } case V_008F0C_BUF_DATA_FORMAT_10_11_11: switch (nfmt) { default: UNREACHABLE("bad nfmt"); case V_008F0C_BUF_NUM_FORMAT_FLOAT: return V_008F0C_GFX11_FORMAT_10_11_11_FLOAT; } } } else if (gfx_level >= GFX10) { unsigned format; switch (dfmt) { default: UNREACHABLE("bad dfmt"); case V_008F0C_BUF_DATA_FORMAT_INVALID: format = V_008F0C_GFX10_FORMAT_INVALID; break; case V_008F0C_BUF_DATA_FORMAT_8: format = V_008F0C_GFX10_FORMAT_8_UINT; break; case V_008F0C_BUF_DATA_FORMAT_8_8: format = V_008F0C_GFX10_FORMAT_8_8_UINT; break; case V_008F0C_BUF_DATA_FORMAT_8_8_8_8: format = V_008F0C_GFX10_FORMAT_8_8_8_8_UINT; break; case V_008F0C_BUF_DATA_FORMAT_16: format = V_008F0C_GFX10_FORMAT_16_UINT; break; case V_008F0C_BUF_DATA_FORMAT_16_16: format = V_008F0C_GFX10_FORMAT_16_16_UINT; break; case V_008F0C_BUF_DATA_FORMAT_16_16_16_16: format = V_008F0C_GFX10_FORMAT_16_16_16_16_UINT; break; case V_008F0C_BUF_DATA_FORMAT_32: format = V_008F0C_GFX10_FORMAT_32_UINT; break; case V_008F0C_BUF_DATA_FORMAT_32_32: format = V_008F0C_GFX10_FORMAT_32_32_UINT; break; case V_008F0C_BUF_DATA_FORMAT_32_32_32: format = V_008F0C_GFX10_FORMAT_32_32_32_UINT; break; case V_008F0C_BUF_DATA_FORMAT_32_32_32_32: format = V_008F0C_GFX10_FORMAT_32_32_32_32_UINT; break; case V_008F0C_BUF_DATA_FORMAT_2_10_10_10: format = V_008F0C_GFX10_FORMAT_2_10_10_10_UINT; break; case V_008F0C_BUF_DATA_FORMAT_10_11_11: format = V_008F0C_GFX10_FORMAT_10_11_11_UINT; break; } // Use the regularity properties of the combined format enum. // // Note: float is incompatible with 8-bit data formats, // [us]{norm,scaled} are incompatible with 32-bit data formats. // [us]scaled are not writable. switch (nfmt) { case V_008F0C_BUF_NUM_FORMAT_UNORM: format -= 4; break; case V_008F0C_BUF_NUM_FORMAT_SNORM: format -= 3; break; case V_008F0C_BUF_NUM_FORMAT_USCALED: format -= 2; break; case V_008F0C_BUF_NUM_FORMAT_SSCALED: format -= 1; break; default: UNREACHABLE("bad nfmt"); case V_008F0C_BUF_NUM_FORMAT_UINT: break; case V_008F0C_BUF_NUM_FORMAT_SINT: format += 1; break; case V_008F0C_BUF_NUM_FORMAT_FLOAT: format += 2; break; } return format; } else { return dfmt | (nfmt << 4); } } #define DUP2(v) v, v #define DUP3(v) v, v, v #define DUP4(v) v, v, v, v #define FMT(dfmt, nfmt) 0xb, {HW_FMT(dfmt, nfmt), HW_FMT(dfmt##_##dfmt, nfmt), HW_FMT_INVALID, HW_FMT(dfmt##_##dfmt##_##dfmt##_##dfmt, nfmt)} #define FMT_32(nfmt) 0xf, {HW_FMT(32, nfmt), HW_FMT(32_32, nfmt), HW_FMT(32_32_32, nfmt), HW_FMT(32_32_32_32, nfmt)} #define FMT_64(nfmt) 0x3, {HW_FMT(32_32, nfmt), HW_FMT(32_32_32_32, nfmt), DUP2(HW_FMT_INVALID)} #define FMTP(dfmt, nfmt) 0xf, {DUP4(HW_FMT(dfmt, nfmt))} #define DST_SEL(x, y, z, w) \ (S_008F0C_DST_SEL_X(V_008F0C_SQ_SEL_##x) | S_008F0C_DST_SEL_Y(V_008F0C_SQ_SEL_##y) | \ S_008F0C_DST_SEL_Z(V_008F0C_SQ_SEL_##z) | S_008F0C_DST_SEL_W(V_008F0C_SQ_SEL_##w)) #define LIST_NFMT_8_16(nfmt) \ [(int)PIPE_FORMAT_R8_##nfmt] = {DST_SEL(X,0,0,1), 1, 1, 1, FMT(8, nfmt)}, \ [(int)PIPE_FORMAT_R8G8_##nfmt] = {DST_SEL(X,Y,0,1), 2, 2, 1, FMT(8, nfmt)}, \ [(int)PIPE_FORMAT_R8G8B8_##nfmt] = {DST_SEL(X,Y,Z,1), 3, 3, 1, FMT(8, nfmt)}, \ [(int)PIPE_FORMAT_B8G8R8_##nfmt] = {DST_SEL(Z,Y,X,1), 3, 3, 1, FMT(8, nfmt)}, \ [(int)PIPE_FORMAT_R8G8B8A8_##nfmt] = {DST_SEL(X,Y,Z,W), 4, 4, 1, FMT(8, nfmt)}, \ [(int)PIPE_FORMAT_B8G8R8A8_##nfmt] = {DST_SEL(Z,Y,X,W), 4, 4, 1, FMT(8, nfmt)}, \ [(int)PIPE_FORMAT_R16_##nfmt] = {DST_SEL(X,0,0,1), 2, 1, 2, FMT(16, nfmt)}, \ [(int)PIPE_FORMAT_R16G16_##nfmt] = {DST_SEL(X,Y,0,1), 4, 2, 2, FMT(16, nfmt)}, \ [(int)PIPE_FORMAT_R16G16B16_##nfmt] = {DST_SEL(X,Y,Z,1), 6, 3, 2, FMT(16, nfmt)}, \ [(int)PIPE_FORMAT_R16G16B16A16_##nfmt] = {DST_SEL(X,Y,Z,W), 8, 4, 2, FMT(16, nfmt)}, #define LIST_NFMT_32_64(nfmt) \ [(int)PIPE_FORMAT_R32_##nfmt] = {DST_SEL(X,0,0,1), 4, 1, 4, FMT_32(nfmt)}, \ [(int)PIPE_FORMAT_R32G32_##nfmt] = {DST_SEL(X,Y,0,1), 8, 2, 4, FMT_32(nfmt)}, \ [(int)PIPE_FORMAT_R32G32B32_##nfmt] = {DST_SEL(X,Y,Z,1), 12, 3, 4, FMT_32(nfmt)}, \ [(int)PIPE_FORMAT_R32G32B32A32_##nfmt] = {DST_SEL(X,Y,Z,W), 16, 4, 4, FMT_32(nfmt)}, \ [(int)PIPE_FORMAT_R64_##nfmt] = {DST_SEL(X,Y,0,0), 8, 1, 8, FMT_64(nfmt)}, \ [(int)PIPE_FORMAT_R64G64_##nfmt] = {DST_SEL(X,Y,Z,W), 16, 2, 8, FMT_64(nfmt)}, \ [(int)PIPE_FORMAT_R64G64B64_##nfmt] = {DST_SEL(X,Y,Z,W), 24, 3, 8, FMT_64(nfmt)}, \ [(int)PIPE_FORMAT_R64G64B64A64_##nfmt] = {DST_SEL(X,Y,Z,W), 32, 4, 8, FMT_64(nfmt)}, \ #define VB_FORMATS \ [(int)PIPE_FORMAT_NONE] = {DST_SEL(0,0,0,1), 0, 4, 0, 0xf, {DUP4(HW_FMT_INVALID)}}, \ LIST_NFMT_8_16(UNORM) \ LIST_NFMT_8_16(SNORM) \ LIST_NFMT_8_16(USCALED) \ LIST_NFMT_8_16(SSCALED) \ LIST_NFMT_8_16(UINT) \ LIST_NFMT_8_16(SINT) \ LIST_NFMT_32_64(UINT) \ LIST_NFMT_32_64(SINT) \ LIST_NFMT_32_64(FLOAT) \ [(int)PIPE_FORMAT_R16_FLOAT] = {DST_SEL(X,0,0,1), 2, 1, 2, FMT(16, FLOAT)}, \ [(int)PIPE_FORMAT_R16G16_FLOAT] = {DST_SEL(X,Y,0,1), 4, 2, 2, FMT(16, FLOAT)}, \ [(int)PIPE_FORMAT_R16G16B16_FLOAT] = {DST_SEL(X,Y,Z,1), 6, 3, 2, FMT(16, FLOAT)}, \ [(int)PIPE_FORMAT_R16G16B16A16_FLOAT] = {DST_SEL(X,Y,Z,W), 8, 4, 2, FMT(16, FLOAT)}, \ [(int)PIPE_FORMAT_B10G10R10A2_UNORM] = {DST_SEL(Z,Y,X,W), 4, 4, 0, FMTP(2_10_10_10, UNORM)}, \ [(int)PIPE_FORMAT_B10G10R10A2_SNORM] = {DST_SEL(Z,Y,X,W), 4, 4, 0, FMTP(2_10_10_10, SNORM), \ AA(AC_ALPHA_ADJUST_SNORM)}, \ [(int)PIPE_FORMAT_B10G10R10A2_USCALED] = {DST_SEL(Z,Y,X,W), 4, 4, 0, FMTP(2_10_10_10, USCALED)}, \ [(int)PIPE_FORMAT_B10G10R10A2_SSCALED] = {DST_SEL(Z,Y,X,W), 4, 4, 0, FMTP(2_10_10_10, SSCALED), \ AA(AC_ALPHA_ADJUST_SSCALED)}, \ [(int)PIPE_FORMAT_B10G10R10A2_UINT] = {DST_SEL(Z,Y,X,W), 4, 4, 0, FMTP(2_10_10_10, UINT)}, \ [(int)PIPE_FORMAT_B10G10R10A2_SINT] = {DST_SEL(Z,Y,X,W), 4, 4, 0, FMTP(2_10_10_10, SINT), \ AA(AC_ALPHA_ADJUST_SINT)}, \ [(int)PIPE_FORMAT_R10G10B10A2_UNORM] = {DST_SEL(X,Y,Z,W), 4, 4, 0, FMTP(2_10_10_10, UNORM)}, \ [(int)PIPE_FORMAT_R10G10B10A2_SNORM] = {DST_SEL(X,Y,Z,W), 4, 4, 0, FMTP(2_10_10_10, SNORM), \ AA(AC_ALPHA_ADJUST_SNORM)}, \ [(int)PIPE_FORMAT_R10G10B10A2_USCALED] = {DST_SEL(X,Y,Z,W), 4, 4, 0, FMTP(2_10_10_10, USCALED)}, \ [(int)PIPE_FORMAT_R10G10B10A2_SSCALED] = {DST_SEL(X,Y,Z,W), 4, 4, 0, FMTP(2_10_10_10, SSCALED), \ AA(AC_ALPHA_ADJUST_SSCALED)}, \ [(int)PIPE_FORMAT_R10G10B10A2_UINT] = {DST_SEL(X,Y,Z,W), 4, 4, 0, FMTP(2_10_10_10, UINT)}, \ [(int)PIPE_FORMAT_R10G10B10A2_SINT] = {DST_SEL(X,Y,Z,W), 4, 4, 0, FMTP(2_10_10_10, SINT), \ AA(AC_ALPHA_ADJUST_SINT)}, \ [(int)PIPE_FORMAT_R11G11B10_FLOAT] = {DST_SEL(X,Y,Z,1), 4, 3, 0, FMTP(10_11_11, FLOAT)}, \ #define HW_FMT(dfmt, nfmt) (V_008F0C_BUF_DATA_FORMAT_##dfmt | (V_008F0C_BUF_NUM_FORMAT_##nfmt << 4)) #define HW_FMT_INVALID (V_008F0C_BUF_DATA_FORMAT_INVALID | (V_008F0C_BUF_NUM_FORMAT_UNORM << 4)) #define AA(v) v static const struct ac_vtx_format_info vb_formats_gfx6_alpha_adjust[] = {VB_FORMATS}; #undef AA #define AA(v) AC_ALPHA_ADJUST_NONE static const struct ac_vtx_format_info vb_formats_gfx6[] = {VB_FORMATS}; #undef HW_FMT_INVALID #undef HW_FMT #define HW_FMT(dfmt, nfmt) V_008F0C_GFX10_FORMAT_##dfmt##_##nfmt #define HW_FMT_INVALID V_008F0C_GFX10_FORMAT_INVALID static const struct ac_vtx_format_info vb_formats_gfx10[] = {VB_FORMATS}; #undef HW_FMT_INVALID #undef HW_FMT #define HW_FMT(dfmt, nfmt) V_008F0C_GFX11_FORMAT_##dfmt##_##nfmt #define HW_FMT_INVALID V_008F0C_GFX11_FORMAT_INVALID static const struct ac_vtx_format_info vb_formats_gfx11[] = {VB_FORMATS}; const struct ac_vtx_format_info * ac_get_vtx_format_info_table(enum amd_gfx_level level, bool has_alpha_adjust_bug) { if (level >= GFX11) return vb_formats_gfx11; else if (level >= GFX10) return vb_formats_gfx10; return has_alpha_adjust_bug ? vb_formats_gfx6_alpha_adjust : vb_formats_gfx6; } const struct ac_vtx_format_info * ac_get_vtx_format_info(enum amd_gfx_level level, bool has_alpha_adjust_bug, enum pipe_format fmt) { return &ac_get_vtx_format_info_table(level, has_alpha_adjust_bug)[fmt]; } /** * Check whether the specified fetch size is safe to use with MTBUF. * * Split typed vertex buffer loads when necessary to avoid any * alignment issues that trigger memory violations and eventually a GPU * hang. This can happen if the stride (static or dynamic) is unaligned and * also if the VBO offset is aligned to a scalar (eg. stride is 8 and VBO * offset is 2 for R16G16B16A16_SNORM). */ static bool is_fetch_size_safe(const enum amd_gfx_level gfx_level, const struct ac_vtx_format_info* vtx_info, const unsigned offset, const unsigned alignment, const unsigned channels) { if (!(vtx_info->has_hw_format & BITFIELD_BIT(channels - 1))) return false; unsigned vertex_byte_size = vtx_info->chan_byte_size * channels; return (gfx_level >= GFX7 && gfx_level <= GFX9) || (offset % vertex_byte_size == 0 && MAX2(alignment, 1) % vertex_byte_size == 0); } /** * Gets the number of channels that can be safely fetched by MTBUF (typed buffer load) * instructions without triggering alignment-related issues. */ unsigned ac_get_safe_fetch_size(const enum amd_gfx_level gfx_level, const struct ac_vtx_format_info* vtx_info, const unsigned offset, const unsigned max_channels, const unsigned alignment, const unsigned num_channels) { /* Packed formats can't be split. */ if (!vtx_info->chan_byte_size) return vtx_info->num_channels; /* Early exit if the specified number of channels is fine. */ if (is_fetch_size_safe(gfx_level, vtx_info, offset, alignment, num_channels)) return num_channels; /* First, assume that more load instructions are worse and try using a larger data format. */ unsigned new_channels = num_channels + 1; while (new_channels <= max_channels && !is_fetch_size_safe(gfx_level, vtx_info, offset, alignment, new_channels)) { new_channels++; } /* Found a feasible load size. */ if (new_channels <= max_channels) return new_channels; /* Try decreasing load size (at the cost of more load instructions). */ new_channels = num_channels; while (new_channels > 1 && !is_fetch_size_safe(gfx_level, vtx_info, offset, alignment, new_channels)) { new_channels--; } return new_channels; } enum ac_image_dim ac_get_sampler_dim(enum amd_gfx_level gfx_level, enum glsl_sampler_dim dim, bool is_array) { switch (dim) { case GLSL_SAMPLER_DIM_1D: if (gfx_level == GFX9) return is_array ? ac_image_2darray : ac_image_2d; return is_array ? ac_image_1darray : ac_image_1d; case GLSL_SAMPLER_DIM_2D: case GLSL_SAMPLER_DIM_RECT: case GLSL_SAMPLER_DIM_EXTERNAL: return is_array ? ac_image_2darray : ac_image_2d; case GLSL_SAMPLER_DIM_3D: return ac_image_3d; case GLSL_SAMPLER_DIM_CUBE: return ac_image_cube; case GLSL_SAMPLER_DIM_MS: return is_array ? ac_image_2darraymsaa : ac_image_2dmsaa; case GLSL_SAMPLER_DIM_SUBPASS: return ac_image_2darray; case GLSL_SAMPLER_DIM_SUBPASS_MS: return ac_image_2darraymsaa; default: UNREACHABLE("bad sampler dim"); } } enum ac_image_dim ac_get_image_dim(enum amd_gfx_level gfx_level, enum glsl_sampler_dim sdim, bool is_array) { enum ac_image_dim dim = ac_get_sampler_dim(gfx_level, sdim, is_array); /* Match the resource type set in the descriptor. */ if (dim == ac_image_cube || (gfx_level <= GFX8 && dim == ac_image_3d)) dim = ac_image_2darray; else if (sdim == GLSL_SAMPLER_DIM_2D && !is_array && gfx_level == GFX9) { /* When a single layer of a 3D texture is bound, the shader * will refer to a 2D target, but the descriptor has a 3D type. * Since the HW ignores BASE_ARRAY in this case, we need to * send 3 coordinates. This doesn't hurt when the underlying * texture is non-3D. */ dim = ac_image_3d; } return dim; } unsigned ac_get_fs_input_vgpr_cnt(const struct ac_shader_config *config) { unsigned num_input_vgprs = 0; if (G_0286CC_PERSP_SAMPLE_ENA(config->spi_ps_input_addr)) num_input_vgprs += 2; if (G_0286CC_PERSP_CENTER_ENA(config->spi_ps_input_addr)) num_input_vgprs += 2; if (G_0286CC_PERSP_CENTROID_ENA(config->spi_ps_input_addr)) num_input_vgprs += 2; if (G_0286CC_PERSP_PULL_MODEL_ENA(config->spi_ps_input_addr)) num_input_vgprs += 3; if (G_0286CC_LINEAR_SAMPLE_ENA(config->spi_ps_input_addr)) num_input_vgprs += 2; if (G_0286CC_LINEAR_CENTER_ENA(config->spi_ps_input_addr)) num_input_vgprs += 2; if (G_0286CC_LINEAR_CENTROID_ENA(config->spi_ps_input_addr)) num_input_vgprs += 2; if (G_0286CC_LINE_STIPPLE_TEX_ENA(config->spi_ps_input_addr)) num_input_vgprs += 1; if (G_0286CC_POS_X_FLOAT_ENA(config->spi_ps_input_addr)) num_input_vgprs += 1; if (G_0286CC_POS_Y_FLOAT_ENA(config->spi_ps_input_addr)) num_input_vgprs += 1; if (G_0286CC_POS_Z_FLOAT_ENA(config->spi_ps_input_addr)) num_input_vgprs += 1; if (G_0286CC_POS_W_FLOAT_ENA(config->spi_ps_input_addr)) num_input_vgprs += 1; if (G_0286CC_FRONT_FACE_ENA(config->spi_ps_input_addr)) num_input_vgprs += 1; if (G_0286CC_ANCILLARY_ENA(config->spi_ps_input_addr)) num_input_vgprs += 1; if (G_0286CC_SAMPLE_COVERAGE_ENA(config->spi_ps_input_addr)) num_input_vgprs += 1; if (G_0286CC_POS_FIXED_PT_ENA(config->spi_ps_input_addr)) num_input_vgprs += 1; return num_input_vgprs; } uint16_t ac_get_ps_iter_mask(unsigned ps_iter_samples) { /* The bit pattern matches that used by fixed function fragment * processing. */ switch (ps_iter_samples) { case 1: return 0xff; case 2: return 0x55; case 4: return 0x11; case 8: return 0x01; default: UNREACHABLE("invalid sample count"); } } void ac_choose_spi_color_formats(unsigned format, unsigned swap, unsigned ntype, bool is_depth, bool use_rbplus, struct ac_spi_color_formats *formats) { /* Alpha is needed for alpha-to-coverage. * Blending may be with or without alpha. */ unsigned normal = 0; /* most optimal, may not support blending or export alpha */ unsigned alpha = 0; /* exports alpha, but may not support blending */ unsigned blend = 0; /* supports blending, but may not export alpha */ unsigned blend_alpha = 0; /* least optimal, supports blending and exports alpha */ /* Choose the SPI color formats. These are required values for RB+. * Other chips have multiple choices, though they are not necessarily better. */ switch (format) { case V_028C70_COLOR_5_6_5: case V_028C70_COLOR_1_5_5_5: case V_028C70_COLOR_5_5_5_1: case V_028C70_COLOR_4_4_4_4: case V_028C70_COLOR_10_11_11: case V_028C70_COLOR_11_11_10: case V_028C70_COLOR_5_9_9_9: case V_028C70_COLOR_8: case V_028C70_COLOR_8_8: case V_028C70_COLOR_8_8_8_8: case V_028C70_COLOR_10_10_10_2: case V_028C70_COLOR_2_10_10_10: if (ntype == V_028C70_NUMBER_UINT) alpha = blend = blend_alpha = normal = V_028714_SPI_SHADER_UINT16_ABGR; else if (ntype == V_028C70_NUMBER_SINT) alpha = blend = blend_alpha = normal = V_028714_SPI_SHADER_SINT16_ABGR; else alpha = blend = blend_alpha = normal = V_028714_SPI_SHADER_FP16_ABGR; if (!use_rbplus && format == V_028C70_COLOR_8 && ntype != V_028C70_NUMBER_SRGB && swap == V_028C70_SWAP_STD) /* R */ { /* When RB+ is enabled, R8_UNORM should use FP16_ABGR for 2x * exporting performance. Otherwise, use 32_R to remove useless * instructions needed for 16-bit compressed exports. */ blend = normal = V_028714_SPI_SHADER_32_R; } break; case V_028C70_COLOR_16: case V_028C70_COLOR_16_16: case V_028C70_COLOR_16_16_16_16: if (ntype == V_028C70_NUMBER_UNORM || ntype == V_028C70_NUMBER_SNORM) { /* UNORM16 and SNORM16 don't support blending */ if (ntype == V_028C70_NUMBER_UNORM) normal = alpha = V_028714_SPI_SHADER_UNORM16_ABGR; else normal = alpha = V_028714_SPI_SHADER_SNORM16_ABGR; /* Use 32 bits per channel for blending. */ if (format == V_028C70_COLOR_16) { if (swap == V_028C70_SWAP_STD) { /* R */ blend = V_028714_SPI_SHADER_32_R; blend_alpha = V_028714_SPI_SHADER_32_AR; } else if (swap == V_028C70_SWAP_ALT_REV) /* A */ blend = blend_alpha = V_028714_SPI_SHADER_32_AR; else assert(0); } else if (format == V_028C70_COLOR_16_16) { if (swap == V_028C70_SWAP_STD || swap == V_028C70_SWAP_STD_REV) { /* RG or GR */ blend = V_028714_SPI_SHADER_32_GR; blend_alpha = V_028714_SPI_SHADER_32_ABGR; } else if (swap == V_028C70_SWAP_ALT) /* RA */ blend = blend_alpha = V_028714_SPI_SHADER_32_AR; else assert(0); } else /* 16_16_16_16 */ blend = blend_alpha = V_028714_SPI_SHADER_32_ABGR; } else if (ntype == V_028C70_NUMBER_UINT) alpha = blend = blend_alpha = normal = V_028714_SPI_SHADER_UINT16_ABGR; else if (ntype == V_028C70_NUMBER_SINT) alpha = blend = blend_alpha = normal = V_028714_SPI_SHADER_SINT16_ABGR; else if (ntype == V_028C70_NUMBER_FLOAT) alpha = blend = blend_alpha = normal = V_028714_SPI_SHADER_FP16_ABGR; else assert(0); break; case V_028C70_COLOR_32: if (swap == V_028C70_SWAP_STD) { /* R */ blend = normal = V_028714_SPI_SHADER_32_R; alpha = blend_alpha = V_028714_SPI_SHADER_32_AR; } else if (swap == V_028C70_SWAP_ALT_REV) /* A */ alpha = blend = blend_alpha = normal = V_028714_SPI_SHADER_32_AR; else assert(0); break; case V_028C70_COLOR_32_32: if (swap == V_028C70_SWAP_STD || swap == V_028C70_SWAP_STD_REV) { /* RG or GR */ blend = normal = V_028714_SPI_SHADER_32_GR; alpha = blend_alpha = V_028714_SPI_SHADER_32_ABGR; } else if (swap == V_028C70_SWAP_ALT) /* RA */ alpha = blend = blend_alpha = normal = V_028714_SPI_SHADER_32_AR; else assert(0); break; case V_028C70_COLOR_32_32_32_32: case V_028C70_COLOR_8_24: case V_028C70_COLOR_24_8: case V_028C70_COLOR_X24_8_32_FLOAT: alpha = blend = blend_alpha = normal = V_028714_SPI_SHADER_32_ABGR; break; default: assert(0); return; } /* The DB->CB copy needs 32_ABGR. */ if (is_depth) alpha = blend = blend_alpha = normal = V_028714_SPI_SHADER_32_ABGR; formats->normal = normal; formats->alpha = alpha; formats->blend = blend; formats->blend_alpha = blend_alpha; } void ac_compute_late_alloc(const struct radeon_info *info, bool ngg, bool ngg_culling, bool uses_scratch, unsigned *late_alloc_wave64, unsigned *cu_mask) { *late_alloc_wave64 = 0; /* The limit is per SA. */ *cu_mask = 0xffff; /* This should never be called on gfx12. Gfx12 doesn't need to mask CUs for late alloc. */ assert(info->gfx_level < GFX12); /* CU masking can decrease performance and cause a hang with <= 2 CUs per SA. */ if (info->min_good_cu_per_sa <= 2) return; /* If scratch is used with late alloc, the GPU could deadlock if PS uses scratch too. A more * complicated computation is needed to enable late alloc with scratch (see PAL). */ if (uses_scratch) return; /* Late alloc is not used for NGG on Navi14 due to a hw bug. */ if (ngg && info->family == CHIP_NAVI14) return; if (info->gfx_level >= GFX10) { /* For Wave32, the hw will launch twice the number of late alloc waves, so 1 == 2x wave32. * These limits are estimated because they are all safe but they vary in performance. */ if (ngg_culling) *late_alloc_wave64 = info->min_good_cu_per_sa * 10; else if (info->gfx_level >= GFX11) *late_alloc_wave64 = 63; else *late_alloc_wave64 = info->min_good_cu_per_sa * 4; /* Limit LATE_ALLOC_GS to prevent a hang (hw bug) on gfx10. */ if (info->gfx_level == GFX10 && ngg) *late_alloc_wave64 = MIN2(*late_alloc_wave64, 64); /* Gfx10: CU2 & CU3 must be disabled to prevent a hw deadlock. * Others: CU1 must be disabled to prevent a hw deadlock. * * The deadlock is caused by late alloc, which usually increases performance. */ *cu_mask &= info->gfx_level == GFX10 ? ~BITFIELD_RANGE(2, 2) : ~BITFIELD_RANGE(1, 1); } else { if (info->min_good_cu_per_sa <= 4) { /* Too few available compute units per SA. Disallowing VS to run on one CU could hurt us * more than late VS allocation would help. * * 2 is the highest safe number that allows us to keep all CUs enabled. */ *late_alloc_wave64 = 2; } else { /* This is a good initial value, allowing 1 late_alloc wave per SIMD on num_cu - 2. */ *late_alloc_wave64 = (info->min_good_cu_per_sa - 2) * 4; } /* VS can't execute on one CU if the limit is > 2. */ if (*late_alloc_wave64 > 2) *cu_mask = 0xfffe; /* 1 CU disabled */ } /* Max number that fits into the register field. */ if (ngg) /* GS */ *late_alloc_wave64 = MIN2(*late_alloc_wave64, G_00B204_SPI_SHADER_LATE_ALLOC_GS_GFX10(~0u)); else /* VS */ *late_alloc_wave64 = MIN2(*late_alloc_wave64, G_00B11C_LIMIT(~0u)); } unsigned ac_compute_cs_workgroup_size(const uint16_t sizes[3], bool variable, unsigned max) { if (variable) return max; return sizes[0] * sizes[1] * sizes[2]; } unsigned ac_compute_lshs_workgroup_size(enum amd_gfx_level gfx_level, mesa_shader_stage stage, unsigned tess_num_patches, unsigned tess_patch_in_vtx, unsigned tess_patch_out_vtx) { /* When tessellation is used, API VS runs on HW LS, API TCS runs on HW HS. * These two HW stages are merged on GFX9+. */ bool merged_shaders = gfx_level >= GFX9; unsigned ls_workgroup_size = tess_num_patches * tess_patch_in_vtx; unsigned hs_workgroup_size = tess_num_patches * tess_patch_out_vtx; if (merged_shaders) return MAX2(ls_workgroup_size, hs_workgroup_size); else if (stage == MESA_SHADER_VERTEX) return ls_workgroup_size; else if (stage == MESA_SHADER_TESS_CTRL) return hs_workgroup_size; else UNREACHABLE("invalid LSHS shader stage"); } unsigned ac_compute_ngg_workgroup_size(unsigned es_verts, unsigned gs_inst_prims, unsigned max_vtx_out, unsigned prim_amp_factor) { /* NGG always operates in workgroups. * * For API VS/TES/GS: * - 1 invocation per input vertex * - 1 invocation per input primitive * * The same invocation can process both an input vertex and primitive, * however 1 invocation can only output up to 1 vertex and 1 primitive. */ unsigned max_vtx_in = es_verts < 256 ? es_verts : 3 * gs_inst_prims; unsigned max_prim_in = gs_inst_prims; unsigned max_prim_out = gs_inst_prims * prim_amp_factor; unsigned workgroup_size = MAX4(max_vtx_in, max_vtx_out, max_prim_in, max_prim_out); return CLAMP(workgroup_size, 1, 256); } static unsigned get_tcs_wg_output_mem_size(uint32_t num_tcs_output_cp, uint32_t num_mem_tcs_outputs, uint32_t num_mem_tcs_patch_outputs, uint32_t num_patches) { /* Align each per-vertex and per-patch output to 16 vec4 elements = 256B. It's most optimal when * the 16 vec4 elements are written by 16 consecutive lanes. * * 256B is the granularity of interleaving memory channels, which means a single output store * in wave64 will cover 4 channels (1024B). If an output was only aligned to 128B, wave64 could * cover 5 channels (128B .. 1.125K) instead of 4, which could increase VMEM latency. */ unsigned mem_one_pervertex_output = align(16 * num_tcs_output_cp * num_patches, AMD_MEMCHANNEL_INTERLEAVE_BYTES); unsigned mem_one_perpatch_output = align(16 * num_patches, AMD_MEMCHANNEL_INTERLEAVE_BYTES); return mem_one_pervertex_output * num_mem_tcs_outputs + mem_one_perpatch_output * num_mem_tcs_patch_outputs; } uint32_t ac_compute_num_tess_patches(const struct ac_compiler_info *info, uint32_t num_tcs_input_cp, uint32_t num_tcs_output_cp, uint32_t num_mem_tcs_outputs, uint32_t num_mem_tcs_patch_outputs, uint32_t lds_per_patch, uint32_t wave_size, bool tess_uses_primid) { /* The VGT HS block increments the patch ID unconditionally within a single threadgroup. * This results in incorrect patch IDs when instanced draws are used. * * The intended solution is to restrict threadgroups to a single instance by setting * SWITCH_ON_EOI, which should cause IA to split instances up. However, this doesn't work * correctly on GFX6 when there is no other SE to switch to. */ if (info->has_primid_instancing_bug && tess_uses_primid) return 1; /* 256 threads per workgroup is the hw limit, but 192 performs better. */ const unsigned num_threads_per_patch = MAX2(num_tcs_input_cp, num_tcs_output_cp); unsigned num_patches = 192 / num_threads_per_patch; /* 127 is the maximum value that fits in tcs_offchip_layout. */ num_patches = MIN2(num_patches, 127); /* When distributed tessellation is unsupported, switch between SEs * at a higher frequency to manually balance the workload between SEs. */ if (info->smaller_tcs_workgroups) num_patches = MIN2(num_patches, 16); /* recommended */ /* Make sure the output data fits in the offchip buffer */ unsigned mem_size = get_tcs_wg_output_mem_size(num_tcs_output_cp, num_mem_tcs_outputs, num_mem_tcs_patch_outputs, num_patches); if (mem_size > info->hs_offchip_workgroup_dw_size * 4) { /* Find the number of patches that fit in memory. Each output is aligned separately, * so this division won't return a precise result. */ num_patches = info->hs_offchip_workgroup_dw_size * 4 / get_tcs_wg_output_mem_size(num_tcs_output_cp, num_mem_tcs_outputs, num_mem_tcs_patch_outputs, 1); assert(get_tcs_wg_output_mem_size(num_tcs_output_cp, num_mem_tcs_outputs, num_mem_tcs_patch_outputs, num_patches) <= info->hs_offchip_workgroup_dw_size * 4); while (get_tcs_wg_output_mem_size(num_tcs_output_cp, num_mem_tcs_outputs, num_mem_tcs_patch_outputs, num_patches + 1) <= info->hs_offchip_workgroup_dw_size * 4) num_patches++; } /* Make sure that the data fits in LDS. This assumes the shaders only * use LDS for the inputs and outputs. */ if (lds_per_patch) { /* LS/HS can only access up to 32K on GFX6-8 and 64K on GFX9+. * * 32K performs the best. We could use 64K on GFX9+, but it doesn't perform well because * 64K prevents GS and PS from running on the same CU. */ const unsigned max_lds_size = 32 * 1024 - AC_TESS_LEVEL_VOTE_LDS_BYTES; num_patches = MIN2(num_patches, max_lds_size / lds_per_patch); assert(num_patches * lds_per_patch <= max_lds_size); } num_patches = MAX2(num_patches, 1); /* Make sure that vector lanes are fully occupied by cutting off the last wave * if it's only partially filled. */ const unsigned threads_per_tg = num_patches * num_threads_per_patch; if (threads_per_tg > wave_size && (wave_size - threads_per_tg % wave_size >= MAX2(num_threads_per_patch, 8))) num_patches = (threads_per_tg & ~(wave_size - 1)) / num_threads_per_patch; if (info->gfx_level == GFX6) { /* GFX6 bug workaround, related to power management. Limit LS-HS * threadgroups to only one wave. */ const unsigned one_wave = wave_size / num_threads_per_patch; num_patches = MIN2(num_patches, one_wave); } /* This is the maximum number that fits into tcs_offchip_layout. */ assert(num_patches <= 127); return num_patches; } uint32_t ac_apply_cu_en(uint32_t value, uint32_t clear_mask, unsigned value_shift, const struct radeon_info *info) { /* Register field position and mask. */ uint32_t cu_en_mask = ~clear_mask; unsigned cu_en_shift = ffs(cu_en_mask) - 1; /* The value being set. */ uint32_t cu_en = (value & cu_en_mask) >> cu_en_shift; uint32_t set_cu_en = info->spi_cu_en; if (info->gfx_level >= GFX12 && clear_mask == 0) { /* The CU mask has 32 bits and is per SE, not per SA. This math doesn't work with * asymmetric WGP harvesting because SA0 doesn't always end on the same bit. */ set_cu_en &= BITFIELD_MASK(info->max_good_cu_per_sa); set_cu_en |= set_cu_en << info->max_good_cu_per_sa; } /* AND the field by spi_cu_en. */ uint32_t spi_cu_en = info->spi_cu_en >> value_shift; return (value & ~cu_en_mask) | (((cu_en & spi_cu_en) << cu_en_shift) & cu_en_mask); } /* Compute the optimal scratch wavesize. */ uint32_t ac_compute_scratch_wavesize(const struct radeon_info *info, uint32_t bytes_per_wave) { /* Add 1 scratch item to make the number of items odd. This should improve * scratch performance by more randomly distributing scratch waves among * memory channels. * * On GFX11+, this is exactly "|= AMD_MEMCHANNEL_INTERLEAVE_BYTES". */ if (bytes_per_wave) bytes_per_wave |= info->scratch_wavesize_granularity; return bytes_per_wave; } /* Return the scratch register value. */ void ac_get_scratch_tmpring_size(const struct radeon_info *info, unsigned num_scratch_waves, unsigned bytes_per_wave, uint32_t *tmpring_size) { /* SPI_TMPRING_SIZE and COMPUTE_TMPRING_SIZE are essentially scratch buffer descriptors. * WAVES means NUM_RECORDS. WAVESIZE is the size of each element, meaning STRIDE. * Thus, WAVESIZE must be constant while the scratch buffer is being used by the GPU. * * If you want to increase WAVESIZE without waiting for idle, you need to allocate a new * scratch buffer and use it instead. This will result in multiple scratch buffers being * used at the same time, each with a different WAVESIZE. * * If you want to decrease WAVESIZE, you don't have to. There is no advantage in decreasing * WAVESIZE after it's been increased. * * Shaders with SCRATCH_EN=0 don't allocate scratch space. */ /* The compiler shader backend should be reporting aligned scratch_sizes. */ assert((bytes_per_wave & BITFIELD_MASK(info->scratch_wavesize_granularity_shift)) == 0 && "scratch size per wave should be aligned"); if (info->gfx_level >= GFX11) num_scratch_waves /= info->max_se; /* WAVES is per SE */ *tmpring_size = S_0286E8_WAVES(num_scratch_waves) | S_0286E8_WAVESIZE(bytes_per_wave >> info->scratch_wavesize_granularity_shift); } /* Convert chip-agnostic memory access flags into hw-specific cache flags. */ union ac_hw_cache_flags ac_get_hw_cache_flags(enum amd_gfx_level gfx_level, enum gl_access_qualifier access, enum ac_access_type type) { union ac_hw_cache_flags result; result.value = 0; bool is_store = type == ac_access_type_store || type == ac_access_type_store_subdword; assert(!(access & ACCESS_SMEM_AMD) || type == ac_access_type_load); assert(!(access & ACCESS_IS_SWIZZLED_AMD) || !(access & ACCESS_SMEM_AMD)); bool scope_is_device = access & (ACCESS_COHERENT | ACCESS_VOLATILE); if (gfx_level >= GFX12) { if (access & ACCESS_CP_GE_COHERENT_AMD) { bool cp_sdma_ge_use_system_memory_scope = gfx_level == GFX12; result.gfx12.scope = cp_sdma_ge_use_system_memory_scope ? gfx12_scope_memory : gfx12_scope_device; } else if (scope_is_device) { result.gfx12.scope = gfx12_scope_device; } else { result.gfx12.scope = gfx12_scope_cu; } if (access & ACCESS_NON_TEMPORAL) { if (type == ac_access_type_load) { /* Don't use non_temporal for SMEM because it can't set regular_temporal for MALL. */ if (!(access & ACCESS_SMEM_AMD)) result.gfx12.temporal_hint = gfx12_load_near_non_temporal_far_regular_temporal; } else if (is_store) { result.gfx12.temporal_hint = gfx12_store_near_non_temporal_far_regular_temporal; } else { result.gfx12.temporal_hint = gfx12_atomic_non_temporal; } } } else if (gfx_level >= GFX11) { /* GFX11 simplified it and exposes what is actually useful. * * GLC means device scope for loads only. (stores and atomics are always device scope) * SLC means non-temporal for GL1 and GL2 caches. (GL1 = hit-evict, GL2 = stream, unavailable in SMEM) * DLC means non-temporal for MALL. (noalloc, i.e. coherent bypass) * * GL0 doesn't have a non-temporal flag, so you always get LRU caching in CU scope. */ if (type == ac_access_type_load && scope_is_device) result.value |= ac_glc; if (access & ACCESS_NON_TEMPORAL && !(access & ACCESS_SMEM_AMD)) result.value |= ac_slc; } else if (gfx_level >= GFX10) { /* GFX10-10.3: * * VMEM and SMEM loads (SMEM only supports the first four): * !GLC && !DLC && !SLC means CU scope <== use for normal loads with CU scope * GLC && !DLC && !SLC means SA scope * !GLC && DLC && !SLC means CU scope, GL1 bypass * GLC && DLC && !SLC means device scope <== use for normal loads with device scope * !GLC && !DLC && SLC means CU scope, non-temporal (GL0 = GL1 = hit-evict, GL2 = stream) <== use for non-temporal loads with CU scope * GLC && !DLC && SLC means SA scope, non-temporal (GL1 = hit-evict, GL2 = stream) * !GLC && DLC && SLC means CU scope, GL0 non-temporal, GL1-GL2 coherent bypass (GL0 = hit-evict, GL1 = bypass, GL2 = noalloc) * GLC && DLC && SLC means device scope, GL2 coherent bypass (noalloc) <== use for non-temporal loads with device scope * * VMEM stores/atomics (stores are CU scope only if they overwrite the whole cache line, * atomics are always device scope, GL1 is always bypassed): * !GLC && !DLC && !SLC means CU scope <== use for normal stores with CU scope * GLC && !DLC && !SLC means device scope <== use for normal stores with device scope * !GLC && DLC && !SLC means CU scope, GL2 non-coherent bypass * GLC && DLC && !SLC means device scope, GL2 non-coherent bypass * !GLC && !DLC && SLC means CU scope, GL2 non-temporal (stream) <== use for non-temporal stores with CU scope * GLC && !DLC && SLC means device scope, GL2 non-temporal (stream) <== use for non-temporal stores with device scope * !GLC && DLC && SLC means CU scope, GL2 coherent bypass (noalloc) * GLC && DLC && SLC means device scope, GL2 coherent bypass (noalloc) * * "stream" allows write combining in GL2. "coherent bypass" doesn't. * "non-coherent bypass" doesn't guarantee ordering with any coherent stores. */ if (scope_is_device && type != ac_access_type_atomic) result.value |= ac_glc | (type == ac_access_type_load ? ac_dlc : 0); if (access & ACCESS_NON_TEMPORAL && !(access & ACCESS_SMEM_AMD)) result.value |= ac_slc; } else { /* GFX6-GFX9: * * VMEM loads: * !GLC && !SLC means CU scope * GLC && !SLC means (GFX6: device scope, GFX7-9: device scope [*]) * !GLC && SLC means (GFX6: CU scope, GFX7: device scope, GFX8-9: CU scope), GL2 non-temporal (stream) * GLC && SLC means device scope, GL2 non-temporal (stream) * * VMEM stores (atomics don't have [*]): * !GLC && !SLC means (GFX6: CU scope, GFX7-9: device scope [*]) * GLC && !SLC means (GFX6-7: device scope, GFX8-9: device scope [*]) * !GLC && SLC means (GFX6: CU scope, GFX7-9: device scope [*]), GL2 non-temporal (stream) * GLC && SLC means device scope, GL2 non-temporal (stream) * * [*] data can be cached in GL1 for future CU scope * * SMEM loads: * GLC means device scope (available on GFX8+) */ if (scope_is_device && type != ac_access_type_atomic) { /* SMEM doesn't support the device scope on GFX6-7. */ assert(gfx_level >= GFX8 || !(access & ACCESS_SMEM_AMD)); result.value |= ac_glc; } if (access & ACCESS_NON_TEMPORAL && !(access & ACCESS_SMEM_AMD)) result.value |= ac_slc; /* GFX6 has a TC L1 bug causing corruption of 8bit/16bit stores. All store opcodes not * aligned to a dword are affected. */ if (gfx_level == GFX6 && type == ac_access_type_store_subdword) result.value |= ac_glc; } if (access & ACCESS_IS_SWIZZLED_AMD) { if (gfx_level >= GFX12) result.gfx12.swizzled = true; else result.value |= ac_swizzled; } return result; } unsigned ac_get_all_edge_flag_bits(enum amd_gfx_level gfx_level) { return gfx_level >= GFX12 ? ((1u << 8) | (1u << 17) | (1u << 26)) : ((1u << 9) | (1u << 19) | (1u << 29)); } /** * Returns a unique index for a per-patch semantic name and index. The index * must be less than 32, so that a 32-bit bitmask of used inputs or outputs * can be calculated. */ unsigned ac_shader_io_get_unique_index_patch(unsigned semantic) { switch (semantic) { case VARYING_SLOT_TESS_LEVEL_OUTER: return 0; case VARYING_SLOT_TESS_LEVEL_INNER: return 1; default: if (semantic >= VARYING_SLOT_PATCH0 && semantic < VARYING_SLOT_PATCH0 + 30) return 2 + (semantic - VARYING_SLOT_PATCH0); assert(!"invalid semantic"); return 0; } } static void clamp_gsprims_to_esverts(unsigned *max_gsprims, unsigned max_esverts, unsigned min_verts_per_prim, bool use_adjacency) { unsigned max_reuse = max_esverts - min_verts_per_prim; if (use_adjacency) max_reuse /= 2; *max_gsprims = MIN2(*max_gsprims, 1 + max_reuse); } void ac_legacy_gs_compute_subgroup_info(enum mesa_prim input_prim, unsigned gs_vertices_out, unsigned gs_invocations, unsigned esgs_vertex_stride, ac_legacy_gs_subgroup_info *out) { unsigned gs_num_invocations = MAX2(gs_invocations, 1); bool uses_adjacency = mesa_prim_has_adjacency((enum mesa_prim)input_prim); const unsigned max_verts_per_prim = mesa_vertices_per_prim(input_prim); /* All these are in dwords: */ /* We can't allow using the whole LDS, because GS waves compete with * other shader stages for LDS space. */ const unsigned max_lds_size = 8 * 1024; const unsigned esgs_itemsize = esgs_vertex_stride / 4; unsigned esgs_lds_size; /* All these are per subgroup: */ const unsigned max_out_prims = 32 * 1024; const unsigned max_es_verts = 255; const unsigned ideal_gs_prims = 64; unsigned max_gs_prims, gs_prims; unsigned min_es_verts, es_verts, worst_case_es_verts; if (uses_adjacency || gs_num_invocations > 1) max_gs_prims = 127 / gs_num_invocations; else max_gs_prims = 255; /* MAX_PRIMS_PER_SUBGROUP = gs_prims * max_vert_out * gs_invocations. * Make sure we don't go over the maximum value. */ if (gs_vertices_out > 0) { max_gs_prims = MIN2(max_gs_prims, max_out_prims / (gs_vertices_out * gs_num_invocations)); } assert(max_gs_prims > 0); /* If the primitive has adjacency, halve the number of vertices * that will be reused in multiple primitives. */ min_es_verts = max_verts_per_prim / (uses_adjacency ? 2 : 1); gs_prims = MIN2(ideal_gs_prims, max_gs_prims); worst_case_es_verts = MIN2(min_es_verts * gs_prims, max_es_verts); /* Compute ESGS LDS size based on the worst case number of ES vertices * needed to create the target number of GS prims per subgroup. */ esgs_lds_size = esgs_itemsize * worst_case_es_verts; /* If total LDS usage is too big, refactor partitions based on ratio * of ESGS item sizes. */ if (esgs_lds_size > max_lds_size) { /* Our target GS Prims Per Subgroup was too large. Calculate * the maximum number of GS Prims Per Subgroup that will fit * into LDS, capped by the maximum that the hardware can support. */ gs_prims = MIN2((max_lds_size / (esgs_itemsize * min_es_verts)), max_gs_prims); assert(gs_prims > 0); worst_case_es_verts = MIN2(min_es_verts * gs_prims, max_es_verts); esgs_lds_size = esgs_itemsize * worst_case_es_verts; assert(esgs_lds_size <= max_lds_size); } /* Now calculate remaining ESGS information. */ if (esgs_lds_size) es_verts = MIN2(esgs_lds_size / esgs_itemsize, max_es_verts); else es_verts = max_es_verts; /* Vertices for adjacency primitives are not always reused, so restore * it for ES_VERTS_PER_SUBGRP. */ min_es_verts = max_verts_per_prim; /* For normal primitives, the VGT only checks if they are past the ES * verts per subgroup after allocating a full GS primitive and if they * are, kick off a new subgroup. But if those additional ES verts are * unique (e.g. not reused) we need to make sure there is enough LDS * space to account for those ES verts beyond ES_VERTS_PER_SUBGRP. */ es_verts -= min_es_verts - 1; out->es_verts_per_subgroup = es_verts; out->gs_prims_per_subgroup = gs_prims; out->gs_inst_prims_in_subgroup = gs_prims * gs_num_invocations; out->max_prims_per_subgroup = out->gs_inst_prims_in_subgroup * gs_vertices_out; out->esgs_lds_size = esgs_lds_size; assert(out->max_prims_per_subgroup <= max_out_prims); } /** * Determine subgroup information like maximum number of vertices and prims. * * This happens before the shader is uploaded, since LDS relocations during * upload depend on the subgroup size. */ bool ac_ngg_compute_subgroup_info(enum amd_gfx_level gfx_level, mesa_shader_stage es_stage, bool is_gs, enum mesa_prim input_prim, unsigned gs_vertices_out, unsigned gs_invocations, unsigned target_workgroup_size, unsigned max_workgroup_size, unsigned wave_size, unsigned esgs_vertex_stride, unsigned ngg_lds_vertex_size, unsigned ngg_lds_scratch_size, bool tess_turns_off_ngg, unsigned max_esgs_lds_padding, ac_ngg_subgroup_info *out) { const unsigned gs_num_invocations = MAX2(gs_invocations, 1); const bool use_adjacency = mesa_prim_has_adjacency(input_prim); const unsigned max_verts_per_prim = mesa_vertices_per_prim(input_prim); const unsigned min_verts_per_prim = is_gs ? max_verts_per_prim : 1; /* All these are in dwords. The maximum is 16K dwords (64KB) of LDS per workgroup. */ /* The LDS scratch is at the beginning of LDS space. */ const unsigned max_lds_size = 16 * 1024 - ngg_lds_scratch_size / 4 - max_esgs_lds_padding / 4; const unsigned target_lds_size = max_lds_size; unsigned esvert_lds_size = 0; unsigned gsprim_lds_size = 0; /* All these are per subgroup: */ const unsigned min_esverts = gfx_level >= GFX11 ? max_verts_per_prim : /* gfx11 requires at least 1 primitive per TG */ gfx_level >= GFX10_3 ? 29 : (24 - 1 + max_verts_per_prim); bool max_vert_out_per_gs_instance = false; unsigned max_gsprims_base, max_esverts_base; /* In the worst case, we can run 1 GS invocation per workgroup. */ assert(!is_gs || gs_vertices_out <= max_workgroup_size); max_gsprims_base = max_esverts_base = target_workgroup_size; if (is_gs) { bool force_multi_cycling = false; unsigned max_out_verts_per_gsprim = gs_vertices_out * gs_num_invocations; retry_select_mode: if (max_out_verts_per_gsprim <= max_workgroup_size && !force_multi_cycling) { if (max_out_verts_per_gsprim) { max_gsprims_base = MIN2(max_gsprims_base, max_workgroup_size / max_out_verts_per_gsprim); } } else { /* Use special multi-cycling mode in which each GS * instance gets its own subgroup. Does not work with * tessellation. */ max_vert_out_per_gs_instance = true; max_gsprims_base = 1; max_out_verts_per_gsprim = gs_vertices_out; } esvert_lds_size = esgs_vertex_stride / 4; gsprim_lds_size = (ngg_lds_vertex_size / 4) * max_out_verts_per_gsprim; if (gsprim_lds_size > target_lds_size && !force_multi_cycling) { if (tess_turns_off_ngg || es_stage != MESA_SHADER_TESS_EVAL) { force_multi_cycling = true; goto retry_select_mode; } } } else { /* VS and TES. */ esvert_lds_size = ngg_lds_vertex_size / 4; } unsigned max_gsprims = max_gsprims_base; unsigned max_esverts = max_esverts_base; if (esvert_lds_size) max_esverts = MIN2(max_esverts, target_lds_size / esvert_lds_size); if (gsprim_lds_size) max_gsprims = MIN2(max_gsprims, target_lds_size / gsprim_lds_size); max_esverts = MIN2(max_esverts, max_gsprims * max_verts_per_prim); clamp_gsprims_to_esverts(&max_gsprims, max_esverts, min_verts_per_prim, use_adjacency); assert(max_esverts >= max_verts_per_prim && max_gsprims >= 1); if (esvert_lds_size || gsprim_lds_size) { /* Now that we have a rough proportionality between esverts * and gsprims based on the primitive type, scale both of them * down simultaneously based on required LDS space. * * We could be smarter about this if we knew how much vertex * reuse to expect. */ unsigned lds_total = max_esverts * esvert_lds_size + max_gsprims * gsprim_lds_size; if (lds_total > target_lds_size) { max_esverts = max_esverts * target_lds_size / lds_total; max_gsprims = max_gsprims * target_lds_size / lds_total; max_esverts = MIN2(max_esverts, max_gsprims * max_verts_per_prim); clamp_gsprims_to_esverts(&max_gsprims, max_esverts, min_verts_per_prim, use_adjacency); assert(max_esverts >= max_verts_per_prim && max_gsprims >= 1); } } /* Round up towards full wave sizes for better ALU utilization. */ if (!max_vert_out_per_gs_instance) { unsigned orig_max_esverts; unsigned orig_max_gsprims; do { orig_max_esverts = max_esverts; orig_max_gsprims = max_gsprims; max_esverts = align(max_esverts, wave_size); max_esverts = MIN2(max_esverts, max_esverts_base); if (esvert_lds_size) max_esverts = MIN2(max_esverts, (max_lds_size - max_gsprims * gsprim_lds_size) / esvert_lds_size); max_esverts = MIN2(max_esverts, max_gsprims * max_verts_per_prim); /* Hardware restriction: minimum value of max_esverts */ max_esverts = MAX2(max_esverts, min_esverts); max_gsprims = align(max_gsprims, wave_size); max_gsprims = MIN2(max_gsprims, max_gsprims_base); if (gsprim_lds_size) { /* Don't count unusable vertices to the LDS size. Those are vertices above * the maximum number of vertices that can occur in the workgroup, * which is e.g. max_gsprims * 3 for triangles. */ unsigned usable_esverts = MIN2(max_esverts, max_gsprims * max_verts_per_prim); max_gsprims = MIN2(max_gsprims, (max_lds_size - usable_esverts * esvert_lds_size) / gsprim_lds_size); } clamp_gsprims_to_esverts(&max_gsprims, max_esverts, min_verts_per_prim, use_adjacency); assert(max_esverts >= max_verts_per_prim && max_gsprims >= 1); } while (orig_max_esverts != max_esverts || orig_max_gsprims != max_gsprims); /* Verify the restriction. */ assert(max_esverts >= min_esverts); } else { max_esverts = MAX2(max_esverts, min_esverts); } unsigned max_out_vertices = max_vert_out_per_gs_instance ? gs_vertices_out : is_gs ? max_gsprims * gs_num_invocations * gs_vertices_out : max_esverts; assert(max_out_vertices <= max_workgroup_size); out->hw_max_esverts = max_esverts; out->max_gsprims = max_gsprims; out->max_out_verts = max_out_vertices; out->max_vert_out_per_gs_instance = max_vert_out_per_gs_instance; /* Don't count unusable vertices. */ out->esgs_lds_size = MIN2(max_esverts, max_gsprims * max_verts_per_prim) * esvert_lds_size; out->ngg_out_lds_size = max_gsprims * gsprim_lds_size; if (is_gs) out->ngg_out_lds_size += ngg_lds_scratch_size / 4; else out->esgs_lds_size += ngg_lds_scratch_size / 4; assert(out->hw_max_esverts >= min_esverts); /* HW limitation */ /* If asserts are disabled, we use the same conditions to return false */ return max_esverts >= max_verts_per_prim && max_gsprims >= 1 && max_out_vertices <= max_workgroup_size && out->hw_max_esverts >= min_esverts; }