From 17058e07469f2dc5b47b4f820bd5a31b7ed9177c Mon Sep 17 00:00:00 2001 From: Keith Whitwell Date: Fri, 2 May 2008 16:02:18 +0200 Subject: [PATCH] tgsi: Implement fast rsqrtf. Not tested, inactive. --- src/gallium/auxiliary/tgsi/exec/tgsi_exec.c | 6 +++ src/gallium/auxiliary/tgsi/exec/tgsi_exec.h | 10 ++++- src/gallium/auxiliary/tgsi/exec/tgsi_sse2.c | 43 +++++++++++++-------- 3 files changed, 40 insertions(+), 19 deletions(-) diff --git a/src/gallium/auxiliary/tgsi/exec/tgsi_exec.c b/src/gallium/auxiliary/tgsi/exec/tgsi_exec.c index 5d5125f7cbe..826b432f09e 100644 --- a/src/gallium/auxiliary/tgsi/exec/tgsi_exec.c +++ b/src/gallium/auxiliary/tgsi/exec/tgsi_exec.c @@ -88,6 +88,10 @@ #define TEMP_OUTPUT_C TGSI_EXEC_TEMP_OUTPUT_C #define TEMP_PRIMITIVE_I TGSI_EXEC_TEMP_PRIMITIVE_I #define TEMP_PRIMITIVE_C TGSI_EXEC_TEMP_PRIMITIVE_C +#define TEMP_3_I TGSI_EXEC_TEMP_THREE_I +#define TEMP_3_C TGSI_EXEC_TEMP_THREE_C +#define TEMP_HALF_I TGSI_EXEC_TEMP_HALF_I +#define TEMP_HALF_C TGSI_EXEC_TEMP_HALF_C #define TEMP_R0 TGSI_EXEC_TEMP_R0 #define FOR_EACH_CHANNEL(CHAN)\ @@ -262,6 +266,8 @@ tgsi_exec_machine_init( mach->Temps[TEMP_2_I].xyzw[TEMP_2_C].f[i] = 2.0f; mach->Temps[TEMP_128_I].xyzw[TEMP_128_C].f[i] = 128.0f; mach->Temps[TEMP_M128_I].xyzw[TEMP_M128_C].f[i] = -128.0f; + mach->Temps[TEMP_3_I].xyzw[TEMP_3_C].f[i] = 3.0f; + mach->Temps[TEMP_HALF_I].xyzw[TEMP_HALF_C].f[i] = 0.5f; } } diff --git a/src/gallium/auxiliary/tgsi/exec/tgsi_exec.h b/src/gallium/auxiliary/tgsi/exec/tgsi_exec.h index 92e2e5e9859..19bd78df3d3 100644 --- a/src/gallium/auxiliary/tgsi/exec/tgsi_exec.h +++ b/src/gallium/auxiliary/tgsi/exec/tgsi_exec.h @@ -133,9 +133,15 @@ struct tgsi_exec_labels #define TGSI_EXEC_TEMP_PRIMITIVE_I 34 #define TGSI_EXEC_TEMP_PRIMITIVE_C 2 -#define TGSI_EXEC_TEMP_R0 35 +#define TGSI_EXEC_TEMP_THREE_I 34 +#define TGSI_EXEC_TEMP_THREE_C 3 -#define TGSI_EXEC_NUM_TEMPS (32 + 4) +#define TGSI_EXEC_TEMP_HALF_I 35 +#define TGSI_EXEC_TEMP_HALF_C 0 + +#define TGSI_EXEC_TEMP_R0 36 + +#define TGSI_EXEC_NUM_TEMPS (32 + 5) #define TGSI_EXEC_NUM_ADDRS 1 #define TGSI_EXEC_NUM_IMMEDIATES 256 diff --git a/src/gallium/auxiliary/tgsi/exec/tgsi_sse2.c b/src/gallium/auxiliary/tgsi/exec/tgsi_sse2.c index 2fd76a3072d..dbf002130bb 100755 --- a/src/gallium/auxiliary/tgsi/exec/tgsi_sse2.c +++ b/src/gallium/auxiliary/tgsi/exec/tgsi_sse2.c @@ -36,7 +36,11 @@ #ifdef PIPE_ARCH_X86 -#define HIGH_PRECISION 1 /* for 1/sqrt() */ +/* for 1/sqrt() + * + * This costs about 100fps (close to 10%) in gears: + */ +#define HIGH_PRECISION 1 #define FOR_EACH_CHANNEL( CHAN )\ @@ -794,20 +798,25 @@ emit_rsqrt( * * See: http://softwarecommunity.intel.com/articles/eng/1818.htm */ - /* This is some code that woudl do the above for a scalar 'a'. We - * obviously are interested in a vector version: - * - * movss xmm3, a; - * movss xmm1, half; - * movss xmm2, three; - * rsqrtss xmm0, xmm3; - * mulss xmm3, xmm0; - * mulss xmm1, xmm0; - * mulss xmm3, xmm0; - * subss xmm2, xmm3; - * mulss xmm1, xmm2; - * movss x, xmm1; - */ + { + struct x86_reg dst = make_xmm( xmm_dst ); + struct x86_reg src = make_xmm( xmm_src ); + struct x86_reg tmp0 = make_xmm( 2 ); + struct x86_reg tmp1 = make_xmm( 3 ); + + assert( xmm_dst != xmm_src ); + assert( xmm_dst != 2 && xmm_dst != 3 ); + assert( xmm_src != 2 && xmm_src != 3 ); + + sse_movaps( func, dst, get_temp( TGSI_EXEC_TEMP_HALF_I, TGSI_EXEC_TEMP_HALF_C ) ); + sse_movaps( func, tmp0, get_temp( TGSI_EXEC_TEMP_THREE_I, TGSI_EXEC_TEMP_THREE_C ) ); + sse_rsqrtps( func, tmp1, src ); + sse_mulps( func, src, tmp1 ); + sse_mulps( func, dst, tmp1 ); + sse_mulps( func, src, tmp1 ); + sse_subps( func, tmp0, src ); + sse_mulps( func, dst, tmp0 ); + } #endif #else /* On Intel CPUs at least, this is only accurate to 12 bits -- not @@ -1295,9 +1304,9 @@ emit_instruction( case TGSI_OPCODE_RSQ: /* TGSI_OPCODE_RECIPSQRT */ FETCH( func, *inst, 0, 0, CHAN_X ); - emit_rsqrt( func, 0, 0 ); + emit_rsqrt( func, 1, 0 ); FOR_EACH_DST0_ENABLED_CHANNEL( *inst, chan_index ) { - STORE( func, *inst, 0, 0, chan_index ); + STORE( func, *inst, 1, 0, chan_index ); } break; -- 2.30.2